The exercises in learning mechanistic interpretability
So, truly important to not rush and also to not just deixar agentes executarem a pesquisa. There is no perspective on this. There is no meaning in that, because there is no learning or investigation. Just models working. Sure, great to understand a great strategy. The goal-driven perspective and the top-venues discuss the learning investigation. Great.
So, do aprendizado com mechanistic interpretability, interesting. The exercises consistiram em identify parameters in the Config object, the cfg, and in understand the process to convert string to tokens and check how the model predicts the next token and with a string, to understand how the model predicted correct. Interesting.
Really great learning.